Person Identification and Authentication with Short-time measured Electroencephalography

Abstract

This thesis proposes a method of distinct person identification and authentication using short-time measured electroencephalography (EEG) designed to have a low error rate. The evaluations are focused on practicality of the measurement process and stability of the connectivity features over time of repeatedly measured EEG at different times along with performances. First, the uniqueness of the features extracted, specifically, power spectrum density (PSD) and z-transformed coherence (ZCOH), are estimated by applying leave-one-out cross validation on data used as the training set. A frequency range of 1-40 Hz is selected from the training results to achieve the best classification accuracy in the testing process. The K-nearest neighbor method is used as the classifier with correlation modified Euclidean distance measurements. Using PSD only, the best correct recognition rate achieved in the 1-40 Hz range is 98.83%, whereas for ZCOH, the best correct recognition rate achieved in the range is 99.67%. Consequently, we conclude that PSD and ZCOH contain uniqueness of individuals. Second, an authentication testing process is conducted with EEG measured over two days and its performance is evaluated. Two types of design modes commonly used in biometrics are applied in this thesis: Registered only mode and Imposter mode. The former mode focuses on whether the test data are well classified to their correct class

therefore, correct recognition rate (CRR) is calculated for overall performance along with the precision, recall, specificity, and F-score for each subject. Consequently, for PSD and ZCOH, averages of 81.35% and 80.04% for precision, 78.20% and 74.88% for recall, 99.04% and 98.81% for specificity, and 0.78 and 0.75 for F-score are achieved at correct recognition rates of 80.39% and 75.70%. The latter, Imposter, mode is designed with a threshold in order to consider the case of other people trying to authenticate. The threshold is selected empirically. To evaluate this mode, false acceptance rate (FAR), false rejection rate (FRR), and half of total error rate (HTER), are calculated. The lowest HTERs achieved for PSD and ZCOH are 13.55% and 19.98%, respectively. Finally, statistical analysis of ZCOH is conducted by inspecting the results of authentication tests in order to filter coherence values that guarantee stability and significance. Filtering combinations of coherence for each subjects frequency bands is achieved by conducting analysis of variance (ANOVA) on data measured on three different days. The results confirm that reliable person identification and authentication with short-time measured EEG using the proposed simple method can be practically adopted and utilized in various applications.